Zachary Rivas

Since the discovery of penicillin in the late 1920s, humans have had an incredibly effective way of treating bacterial infections. However, in the last century, the bacteria have fought back. Bacteria have developed ways to circumvent antibiotics, and now this antibiotic resistance could result in what once would have been considered a minor injury turning deadly because antibiotics no longer work. This is not a distant future; we are already seeing examples of this such as Acinetobacter baumannii, a highly drug-resistant bacteria that thrives in hospital settings and causes severe infections, particularly in immunocompromised patients. Concerningly, antibiotic discovery cannot keep pace with the rapid evolution of resistant bacteria. However, artificial intelligence (AI) may offer a game-changing solution.

In a paper published by Liu et al., the researchers demonstrate how harnessing the power of AI is reshaping drug discovery with unprecedented speed and efficiency¹. The researchers first used known compounds that inhibit A. baumannii growth to train the neural network. The neural network is the model that underpins the AI output. The researchers then applied this neural network model to over 6,500 previously untested molecules. Within 90 minutes, the model identified 240 potential antibiotics. Additional empirical testing by the lab identified the most potent candidate: RS-102895, later renamed Abaucin.

Abaucin was selected for its potent antibacterial properties and for its specificity in targeting A. baumannii. Abaucin has “narrow-spectrum” activity, which means it’s highly effective against A. baumannii but has no or little impact on several other bacterial species. Antibiotic selectivity is crucial because narrow-spectrum agents preserve beneficial microbiota and reduce selective pressure that drives antimicrobial resistance in off-target bacteria. This is particularly important for patients already battling multiple infections, as is often the case with A. baumannii.

While studying Abaucin’s mechanism of action, researchers found that Abaucin specifically targets A. baumannii’s lipoprotein trafficking system. The A. baumannii lipoprotein trafficking system functions differently from that of other bacteria, likely explaining Abaucin’s selectivity. Abaucin inhibits a key protein in the lipoprotein trafficking system, called LolE. By interfering with LolE, Abaucin causes the bacteria to become abnormally large and unable to properly package its cellular contents. This LolE disruption ultimately prevents A. baumannii growth, as initially predicted by the AI model.

After confirming Abaucin’s effectiveness in a dish, the researchers moved to the next step—testing its ability to treat infections in living animals. Liu et al. infected mice with 6.5 million A. baumannii and treated the wounds with a gel, either empty or containing 4% Abaucin. After 25 hours, tissue analysis revealed a stark contrast between the two groups. Mice treated with the empty gel had bacterial levels surge to 690 million bacteria per gram of tissue. In contrast, those treated with the Abaucin-infused gel had 40 million bacteria per gram of tissue—a significant reduction compared to the control. Additionally, mice that received Abaucin treatment exhibited noticeably less inflammation, suggesting both antibacterial efficacy and therapeutic benefits in reducing infection-related damage.

These findings are incredibly promising as a novel method to mine new antibiotics. However, caution should be exercised about Abaucin’s effectiveness. Abaucin didn’t eliminate all A. baumannii in the mice and A. baumannii is infamous for developing resistance from even a small surviving population. Additionally, further in-depth studies on Abaucin’s safety and toxicity in humans are essential before it can be considered for clinical use. Despite these challenges, this study marks a significant milestone for AI’s use to develop medicines in infectious disease research. As the models continue to improve, so will the accuracy and efficiency of drug discovery. More broadly, neural networks like those used in the research done by Liu et al. serve as a proof of concept for AI-driven drug development, with applications beyond antibiotics, including antiviral therapies, autoimmune disease treatments, and cancer therapeutics. This study underscores AI’s potential not as a substitute for human expertise but as a powerful tool to enhance scientific discovery, making researchers more effective, not more replaceable.

1. Liu, G., Catacutan, D.B., Rathod, K. et al. Deep learning-guided discovery of an antibiotic targeting Acinetobacter baumannii. Nat Chem Biol 19, 1342–1350 (2023). https://doi.org/10.1038/s41589-023-01349-8